AI data center investment has exceeded hyperscale cloud providers' own cash flow, with financing expanding from leases to debt, equity, and supply-chain credit support

Overview

The report examines how the next phase of AI data center construction will be funded once hyperscale cloud providers' operating cash flow is insufficient to cover capital expenditures. Its core conclusion is that leases, bonds, equity, customer prepayments, chipmaker support, and SPV project debt collectively expand financing capacity, but all of these structures ultimately represent future cash flow commitments and increase the risks associated with the cost of capital, financial leverage, dilution, and accounting comparability.

Core views

The AI compute buildout has moved beyond the stage of relying solely on internal cash flow. The report expects hyperscale cloud providers' cash capital expenditures to exceed $1.2 trillion in 2027, while their aggregate operating cash flow will be approximately $1 trillion; these companies are reinvesting more than 40% of sales revenue in AI capital expenditures. Even with the high margins of their traditional businesses, this investment intensity cannot be sustained solely by current operating cash flow. Amazon's and Google's free cash flow turned negative in the second quarter of 2026, while Meta's is expected to turn negative in the next quarter. Consequently, incremental capital expenditures are shifting sequentially toward leases, public debt, reduced share repurchases, and equity issuance. The first layer of external financing is off-balance-sheet financing supported by long-term lease and purchase commitments. Once investment-grade cloud providers supply long-term leases, guarantees, or purchase commitments, data center developers and suppliers can use them to borrow from private credit institutions and build capacity before the cloud providers make actual payments or recognize liabilities. Such off-balance-sheet commitments and guarantees disclosed by hyperscale cloud providers, Nvidia, and Broadcom total more than $3.1 trillion. Undiscounted commitments by hyperscale cloud providers alone exceed $2.7 trillion, approximately equivalent to three years of their current operating cash flow, indicating that cash obligations arise earlier and extend for longer than on-balance-sheet leverage suggests. Specifically, hyperscale cloud providers have committed to $1.1 trillion in payments for leases that have not yet commenced. Many core data center buildings are financed by SPVs: the SPVs obtain private credit funding, while investment-grade cloud providers supply credit support through long-term leases. Meanwhile, purchase commitments made by cloud providers, Nvidia, and Broadcom for GPUs, memory, wafer fabrication capacity, networking equipment, and other infrastructure have exceeded $1.7 trillion. The related contracts will gradually flow through free cash flow once facilities are completed, but the growth in the scale and duration of commitments increases operating leverage; if supply and demand normalize earlier than expected, companies may still have to pay for excess capacity or attempt to renegotiate contracts. The second layer of funding comes from on-balance-sheet debt. Once data centers are completed, the related operating or finance lease liabilities begin to enter the balance sheet; in addition, hyperscale cloud providers began issuing bonds in late 2025 and accelerated issuance in 2026. Their aggregate short- and long-term debt and lease liabilities have reached $770 billion. Cloud providers' issuance as a share of nonfinancial investment-grade bond supply increased from 2% in 2025 to 19% year-to-date in 2026. As capital expenditure increases become increasingly dependent on debt rather than cash, project economics depend more heavily on the cost of external capital, and the report also warns that credit spreads on AI financing are widening. Despite the increase in on-balance-sheet liabilities, their scale remains substantially below off-balance-sheet commitments. The degree of alignment between commitments and future revenue also differs across business models. Meta and Google primarily build compute capacity to support internal models, so their commitments exceed remaining performance obligations; for Microsoft and Amazon, which sell compute capacity, commitments are approaching remaining performance obligations. The report uses this distinction to demonstrate that off-balance-sheet commitments should not be viewed merely as forward purchase data, but must also be compared with the contractual revenue or internal returns available to support those commitments. The third layer of funding comes from reducing share repurchases and issuing equity. If interest rates rise, credit spreads widen, or elevated AI investment persists longer than expected, the importance of equity financing may continue to increase. Before undertaking direct equity offerings, cloud providers have already retained cash by reducing share repurchases, causing equity compensation that was previously offset by repurchases to increasingly translate into net equity dilution. Microsoft is the only hyperscale cloud provider that has not yet reduced repurchases or issued debt. Google, by contrast, reduced annual repurchases from more than $60 billion to zero and issued $50 billion of equity in the second quarter of 2026, thereby obtaining more than $110 billion in additional funding for AI infrastructure; after its share count had fallen 13% from its peak over the past decade, shareholders are now beginning to face dilution in the opposite direction. The report expects equity compensation-related dilution to increase as equity compensation per employee continues to grow and companies need to preserve cash for the buildout. New financing sources further bring customer and chip supplier balance sheets into the construction chain. Oracle disclosed $4.6 billion in customer prepayments used for capital expenditures in its most recent quarter. The funds are recorded as deferred revenue, but because they are collected more than a year before revenue recognition, their economic substance is closer to debt with a financing component; interest expense should be recognized in earnings using Oracle's incremental borrowing rate. The report notes that the implied yields to maturity on its 10-year and 30-year bonds are 6.9% and 7.9%, respectively, and that the interest cost of customer prepayments should reflect similar borrowing rates. As an example, the report states that if a customer pays $10 billion two years in advance and the seller's incremental borrowing rate is 10%, the seller would need to recognize approximately $1 billion of interest expense annually, compound it in the second year, and recognize approximately $12 billion of revenue upon delivery. Broadcom and Nvidia help unrated AI labs obtain financing costs close to those of investment-grade suppliers through chip-leasing SPVs. The SPVs issue debt to purchase chips and lease them to AI labs; if lease payments cease and chip resale proceeds are insufficient to repay designated bondholders, the chipmakers provide residual-value support to cover the shortfall. Broadcom has announced chip-leasing arrangements capable of supporting up to 20GW of compute capital expenditures, and Nvidia is also reported to be developing new structures. From an accounting perspective, chip sales revenue may need to be allocated between the chips themselves and the residual-value guarantee, and manufacturers must also recognize a guarantee liability at the fair value that would be charged in a comparable standalone guarantee transaction; if the guarantee ultimately does not require payment, guarantee income generally should not be classified as operating revenue. Finally, the report emphasizes that SPV and lease accounting may obscure economic leverage. Data centers built for and leased to cloud providers generally remain off balance sheet during construction, while the related SPVs may qualify as variable interest entities because the cloud provider is the sole tenant. Consolidation depends on whether the cloud provider both has the power to direct the activities that most significantly affect economic performance and absorbs losses or receives economic benefits. Meta and Google disclose that they do not consolidate because they do not control the key activities, but this judgment requires continuous reassessment; future tenant negotiations, asset sales, operating control, responsibility for cost overruns, or changes in the probability of residual-value guarantees could alter the consolidation conclusion, causing today's off-balance-sheet debt to enter the financial statements in the future. After Microsoft extended the estimated useful life of data centers from 15 years to 25 years, a typical 15-year lease term fell from covering 100% of useful life to 60%, below the 75% finance lease test threshold cited in the report, causing some leases to shift from finance leases to operating leases. Microsoft's disclosed free cash flow includes capital expenditures from finance leases but excludes operating leases, so the new estimate makes reported capital expenditures appear lower and free cash flow appear higher, even though the economic substance of the leases has not changed and still resembles debt-financed construction. When comparing companies' free cash flow, the report recommends adding the value of data centers obtained through long-term operating leases to capital expenditures and free cash flow adjustments, while treating operating lease costs like depreciation and adding them back; in FY2026, Microsoft's lease footnote showed $24.6 billion of assets obtained through finance leases, while Oracle disclosed $18.2 billion of assets obtained through operating leases. Regardless of accounting classification, rating agencies treat lease liabilities as debt. Therefore, the report's overall judgment is not that funding sources have been exhausted, but that the next phase of construction increasingly depends on advance commitments of future cash flow. Leases, SPVs, bonds, equity, and supplier guarantees can extend the buildout cycle, but they do not eliminate financing costs. If incremental commitments push up interest rates or credit spreads, returns on new AI infrastructure investments must cover a higher cost of capital; meanwhile, investors need to look through accounting classifications and jointly assess on-balance-sheet debt, off-balance-sheet commitments, assets obtained through leases, free cash flow adjustments, and equity dilution.

Analysis framework

The report first compares expected 2027 capital expenditures with operating cash flow to confirm the existence of an internal funding gap in the AI buildout; it then analyzes long-term lease and purchase commitments, on-balance-sheet debt, reduced share repurchases and equity issuance, customer prepayments, and chipmaker support in the sequence in which these funding sources emerge. Finally, using VIE consolidation criteria, lease classification, and free cash flow definitions, the report recasts financing arrangements with different legal or accounting forms but similar economic substance as future cash obligations and assesses their effects on leverage, financing costs, profit recognition, and equity dilution.

Methodology notes

• Corporate Fundamentals and Financial FrameworksFree cash flow analysis

  Analysis of operating cash flow, capital expenditures, and free cash flow shortfalls

  The report compares expected capital expenditures with operating cash flow and tracks when Amazon's, Google's, and Meta's free cash flow turns negative to determine when cloud providers need to shift from internal financing to lease, debt, and equity financing.

• Corporate Fundamentals and Financial FrameworksThree-Statement Reconciliation

  Accounting look-through of leases, deferred revenue, guarantee liabilities, and free cash flow

  The report connects the different presentations of financing arrangements across the balance sheet, income statement, and cash flow statement, explaining how operating leases, customer prepayments, and residual-value guarantees affect capital expenditures, interest expense, revenue, and liabilities.

• Fixed Income and Credit AnalysisSpread analysis

  External financing costs and project return thresholds

  Through cloud providers' rising share of bond supply, widening AI financing spreads, and the implied yields to maturity on Oracle's bonds, the report explains that once capital expenditures shift toward debt financing, returns on new projects must cover a higher cost of capital.

• (Out-of-Vocabulary Method)

  VIE primary beneficiary determination and continuous reassessment

  Based on whether cloud providers have decision-making authority over key activities and absorb losses or receive economic benefits, the report determines whether they should consolidate data center SPVs and notes that reassessment must continue throughout the life of the transaction as power and economic risks change.

• (Out-of-Vocabulary Method)

  Comparable adjustment for the economic substance of leases and self-built, debt-financed assets

  The report treats long-term leased data centers and data centers self-built with bond financing as economically similar transactions and uses lease footnotes to adjust capital expenditures and free cash flow to improve comparability across companies.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Microsoft (MSFT)

  As a hyperscale cloud provider and seller of compute capacity, its commitments are approaching remaining performance obligations; changes in data center useful-life estimates have also altered lease classification and free cash flow presentation.

  Strengths

  The report states that it continues to offset dilution through share repurchases and is the only hyperscale cloud provider that has not yet reduced repurchases or issued debt.

  Weaknesses

  Operating leases are not included in the capital expenditures used in its disclosed free cash flow, potentially causing reported capital expenditures to be understated and free cash flow to be overstated.

  Comparison

  Unlike Google and Meta, which have reduced repurchases, Microsoft continues to offset equity dilution.

  Risks

  The economic substance of leases still resembles debt financing, and rating agencies also treat lease liabilities as debt.

• Amazon (AMZN)

  As a hyperscale cloud provider and seller of compute capacity, its construction commitments are approaching remaining performance obligations.

  Strengths

  Compute sales contracts can provide future revenue corresponding to some construction commitments.

  Weaknesses

  Free cash flow turned negative in 2Q26, and internal cash flow no longer fully covers the buildout.

  Comparison

  Compared with Meta and Google, which primarily support internal models, its commitments are better aligned with contractual revenue.

  Risks

  Further increases in capital expenditures will require more lease, debt, or equity funding.

• Google (GOOGL)

  It primarily builds compute capacity for internal models, has commitments exceeding remaining performance obligations, and has freed up funding by halting repurchases and issuing equity.

  Strengths

  It reduced annual repurchases of more than $60 billion to zero and issued $50 billion of equity, freeing up more than $110 billion in construction funding.

  Weaknesses

  Free cash flow turned negative in 2Q26, and net equity dilution has begun to emerge.

  Comparison

  Unlike Microsoft, which continues to use repurchases to offset dilution, Google has shifted substantially toward cash retention and equity financing.

  Risks

  Returns on internal model investments need to cover continuously rising financing and commitment costs.

• Meta (META)

  It primarily builds compute capacity for internal models and uses unconsolidated data center VIE and lease structures.

  Strengths

  Its investment-grade credit can help SPVs obtain private credit funding at a lower cost.

  Weaknesses

  Free cash flow is expected to turn negative next quarter, and commitments supporting internal models exceed remaining performance obligations.

  Comparison

  Compared with sellers of compute capacity such as Microsoft and Amazon, its commitments lack a direct correspondence with contractual revenue of comparable scale.

  Risks

  Changes in VIE power or economic risks could alter the consolidation conclusion, and the related debt could move from off balance sheet to on balance sheet.

• Oracle (ORCL)

  It finances data center capital expenditures through customer prepayments and obtains substantial assets through operating leases.

  Strengths

  It received $4.6 billion in customer prepayments in the most recent quarter, providing an additional source of cash for the buildout.

  Weaknesses

  Long-term prepayments contain a financing component and require recognition of interest expense at the incremental borrowing rate.

  Comparison

  It obtained $18.2 billion of assets through operating leases in FY2026, while Microsoft disclosed $24.6 billion of assets obtained through finance leases.

  Risks

  Interest accretion within deferred revenue will affect future earnings and revenue recognition.

• Nvidia (NVDA) and Broadcom (AVGO)

  By supporting chip-leasing SPVs, they use their own credit and the residual value of chips to help unrated AI labs obtain financing.

  Strengths

  This can enable customers to lease chips at costs closer to those available to investment-grade suppliers and expand available compute financing.

  Weaknesses

  Sales consideration may need to be allocated between the chips and the guarantee, with a guarantee liability recognized at fair value.

  Comparison

  Unlike cloud providers that rely directly on cash, debt, or lease financing, the two companies enhance SPV credit through supplier residual-value support.

  Risks

  If lease payments cease and chip resale values are insufficient, the two companies may need to fulfill their residual-value support obligations.

Key data

• Hyperscale cloud providers' 2027 cash capital expendituresMore than $1.2 trillionAbove expected operating cash flow of approximately $1 trillion.
• AI capital expenditure intensityMore than 40% of sales revenueThe proportion hyperscale cloud providers are reinvesting in AI capital expenditures.
• Free cash flow turning negativeAmazon and Google turned negative in 2Q26; Meta is expected to turn negative next quarterDriving companies toward external funding such as leases, debt, and equity.
• Off-balance-sheet commitments and guaranteesMore than $3.1 trillionThe aggregate amount disclosed by hyperscale cloud providers, NVDA, and AVGO.
• Hyperscale cloud providers' undiscounted commitmentsMore than $2.7 trillionApproximately equivalent to three years of current operating cash flow.
• Lease payment commitments for leases not yet commenced$1.1 trillionPrimarily provides credit support for data center SPV financing.
• Purchase commitmentsMore than $1.7 trillionCovers GPUs, memory, manufacturing capacity, networking equipment, and other items.
• On-balance-sheet debt and lease liabilities$770 billionIncludes hyperscale cloud providers' short- and long-term debt and lease liabilities.
• Share of nonfinancial investment-grade bond supply2% in 2025; 19% year-to-date in 2026Hyperscale cloud providers' share of bond issuance has increased significantly.
• Changes in Google's repurchases and equity financingAnnual repurchases fell from more than $60 billion to zero; $50 billion of equity issued in 2Q26Together, these measures freed up more than $110 billion for AI infrastructure.
• Historical change in Google's share countDown 13% from its peak over the past decadeThe report notes that the previous downward trend has begun to reverse toward shareholder dilution.
• Oracle customer prepayments$4.6 billionDisclosed in the most recent quarter and used to finance capital expenditures.
• Implied yields to maturity on Oracle bonds10-year 6.9%; 30-year 7.9%The report believes the interest cost of customer prepayments should reflect similar borrowing rates.
• Broadcom chip-leasing financing supportUp to 20GWThe announced facility scale for supporting compute capital expenditures.
• Microsoft data center useful-life estimateExtended from 15 years to 25 yearsUseful-life coverage for a typical 15-year lease fell from 100% to 60%, below the 75% test threshold.
• Assets obtained by Microsoft through finance leases$24.6 billionDisclosed in the FY2026 10-K lease footnote.
• Assets obtained by Oracle through operating leases$18.2 billionDisclosed in the FY2026 10-K lease footnote.

Impact & implications

The report believes that the financeable scale of AI infrastructure no longer depends solely on cloud providers' current cash flow, but on their ability to commit future cash through leases, purchase commitments, guarantees, bonds, equity, and customer prepayments. Investors therefore need to assess off-balance-sheet commitments together with on-balance-sheet debt and adjust for differences in capital expenditures and free cash flow caused by operating leases; as the share of external funding, credit spreads, and equity dilution rises, the required returns for new projects to cover the cost of capital will also increase.

Risks

• If interest rates rise or AI financing credit spreads continue to widen, returns on new infrastructure projects may be insufficient to cover the higher cost of capital.
• If compute supply and demand normalize earlier than expected, companies may need to pay for excess capacity or renegotiate long-term lease and purchase commitments.
• Increases in the scale and duration of long-term commitments will raise operating leverage and gradually reduce free cash flow after facilities are completed.
• Reduced share repurchases and increased equity issuance will cause equity compensation to translate more directly into net equity dilution.
• The determination of a VIE's primary beneficiary is not static; changes in control, responsibility for cost overruns, or the probability of residual-value guarantees could trigger consolidation.
• Chip residual-value guarantees may require suppliers to cover shortfalls when lease payments and resale proceeds are insufficient.
• Operating leases can cause reported capital expenditures to be understated and free cash flow to be overstated, thereby understating the economic burden of the buildout.

What to watch

• Track the incremental funding sources used by hyperscale cloud providers when they further increase capital expenditures, as well as their costs.
• Monitor AI financing bond issuance volumes, credit spreads, and changes in interest rates.
• Observe changes in total leases, purchase commitments, and guarantees relative to operating cash flow and remaining performance obligations.
• Track net share-count growth resulting from reductions in cloud provider repurchases, equity issuance, and equity compensation.
• Monitor whether VIE consolidation conclusions for data center SPVs are adjusted because of changes in power or economic risks.
• Observe interest accrual and revenue recognition for Oracle customer prepayments, as well as the accounting treatment of chipmaker SPV guarantee liabilities.
• When comparing companies' disclosed capital expenditures and free cash flow, consider the adjustment impact of assets obtained through operating leases.